Runtime + Decision Log

This page captures the key runtime decisions that are currently active in production.

Last reviewed: 2026-03-10.

Runtime baseline

Nova is natural-language first. The same safety model applies across dashboard, API, and Telegram.

User message
  -> Intent routing
  -> Planning
  -> Validation / repair
  -> Policy gate (auto / approval / blocked)
  -> Execution
  -> Auditing + response

Locked decisions

1. Hierarchical intent routing is always used.
2. One focused clarification is preferred over long back-and-forth.
3. Single repair pass is attempted before asking the user.
4. Confirm-before-write for mutating calendar actions.
5. Natural-language confirmation only auto-runs when exactly one pending approval exists.
6. Replay-safe approvals (especially in Telegram callbacks).

Write safety model

State progression for mutating actions:

proposed -> confirmed -> completed|failed

The goal is to avoid duplicate provider writes even with retries or concurrent confirmations.

Calendar retrieval decision

Nova uses bounded candidate retrieval (deterministic filter first), then ranking. It can use hybrid retrieval when enabled, with deterministic fallback.

Security defaults

• strict ownership checks on approvals
• throttled callback/confirm paths
• safe user-facing errors
• immutable audit events for proposal/confirm/execute lifecycle

Known follow-ups

• continue tuning perf gate thresholds from real CI data
• continue phased rollout of semantic retrieval